融合特征通道重要性与相似性的深度YOLO网络压缩方法

基于深度YOLO网络的目标检测方法网络结构复杂、冗余参数过多、计算量大，极大影响模型检测性能。针对此问题，从降低网络中低效通道和冗余通道的影响出发，提出了一种融合特征通道重要性与相似性的深度YOLO网络压缩方法，基于深度网络压缩中的网络剪枝思路，采用2次剪枝剪去低效及冗余特征通道。构建通道重要性计算方法，将稀疏因子作为通道效能计算指标，结合剪枝率剪去低效通道;根据通道间存在的线性关系计算其相似度，对相似度较高的通道进行替代，剪去相似度较大的通道;微调模型参数，恢复剪枝前的检测精度。仿真实验表明，同当前性能较优的深度网络压缩方案相比，提出的方法在保证目标检测精度的同时极大减小了模型尺寸、提升了检测速度，方法可行、有效。

Compression method combining feature channel importance and similarity for deep YOLO network

Object detection methods based on deep YOLO networks suffer from the problems of complex network structures, redundant parameters, and high computational complexity, which greatly affect the detection performance of the model. Regarding the above issues, we construct a deep YOLO network compression method that integrates feature channel importance and similarity to reduce the impact of inefficient and redundant channels in YOLO network. Based on the network pruning idea in deep network compression, the method uses a two-stage pruning approach to remove inefficient and redundant feature channels. Firstly, a channel importance calculation method is constructed, where sparsity factor is used as an indicator for channel inefficiency, and channels are pruned according to their sorting order and pruning rate. Secondly, the similarity between channels is calculated based on the linear relationship between them, and channels with high similarity can be replaced. After pruning, model parameters are fine-tuned to restore the detection accuracy before pruning. Through simulation experiments on real datasets for object detection, compared with current deep network compression schemes with better performance, the proposed method greatly reduces model size and improves detection speed while ensuring detection accuracy, demonstrating the feasibility and effectiveness of the method.